fix add_special_tokens issue for data with template

[edixiong]

Fix the add_special_tokens issue as is mentioned in this issue: #1400

Fixes #1400

[younesbelkada]

Thanks for the fix ! Can you slightly elaborate on why this should be the right fix? 🙏
cc @philschmid FYI

[edixiong]

Hi @younesbelkada , in line 267, we check if there is a template. If true, dataset_formatting.py will call apply_chat_template and the special token will be added there. In sft_trainer.py, the input will be tokenized and the special token will be added again if we do not explicitly set that add_special_tokens to False. Another workaround is to warn the user to manually set add_special_tokens to False (similar to @philschmid did).

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[younesbelkada]

Thanks !